X local-model bench digest — Oct 31, 2026
As of Oct 31, 2026 — 34 posts collected from X. 87 measured runs catalogued, 1 published to the board, 33 kept to the digest only. Latest source post Oct 31, 2026 09:00 UTC.
Every number below is what its author published, transcribed with the engine, quantization and context length they reported. Sources are linked on each entry. Only posts that state the full methodology (engine build, quantization, context length, hardware state, measurement method) reach the benchmark board; the rest stay here as leads.
DGX Spark × Qwen3.8-Flash-Next
Oct 31, 2026 · by @yuracafe_
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | IQ3_S | strata | 51.3 tok/s | — |
| 16384 | IQ3_S | strata | — | 1940.9 tok/s |
| — | IQ3_S | strata | 77.9 tok/s | — |
| 16384 | IQ3_S | strata | — | 785.6 tok/s |
DGX Spark（GB10）とRX 9070 XTで、Strataのベンチを取ってみました☕ 同じQwen3.8-Flash-Next（IQ3_S／Thinking ON）でも、 ・入力処理と出力開始の早さはGB10 ・生成速度はRX 9070 XT という結果に 短文の生成はGB10が51.3 tok/s、9070 XTが77.9 tok/s 一方、16K入力の処理はGB10が1,940.9 tok/s、9070 XTが785.6 tok/sでした そして、ベンチ取って画像作ってる間にStrataに更新来ました... なので参考程度に... #ローカルLLM #DGXSpark #Radeon
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, hardware state, measurement method).
RTX PRO 5000
Oct 31, 2026 · by @jun1228909
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata v0.2.9 | 53.8 tok/s | — |
| — | — | strata v0.2.9 | 103.7 tok/s | — |
ちょ、0.2.8 持ってきたとこ😂 Strata v0.2.9 リリース 🚀 Linux並列高速化 + Strix Halo強化 + multi-GPU DeepSeek改善 + 1-bit DeepSeek追加😃 ・Linux + NVIDIAのparallel requestを大幅高速化 ・RTX PRO 5000で8並列時、0.2.8比約1.45倍 ・Linuxのprompt処理を標準で約2〜6%高速化 ・2×Arc Pro B65でprompt +4.6〜15.8%、8並列42.1→53.8 tok/s ・Strix Halo + Qwenのpromptを大幅高速化 ・Strix Halo + DeepSeekも最大約1.66倍 ・DGX Spark + DeepSeekのpromptも最大約2倍 ・DeepSeek + 複数AMD GPUで全GPUを自動利用するよう修正 ・4×R9700でresident…
Digest-only: model.name missing; incomplete methodology (quantization, context length, hardware state, measurement method).
DGX Spark × DeepSeek (Unsloth 1-bit, UD-IQ1_S / UD-IQ1_M)
Oct 31, 2026 · by @murasametech
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | strata v0.2.9 | — | 81.0 tok/s |
Strata v0.2.9 リリース！🚀 Unslothの1-bit DeepSeek（UD-IQ1_S / UD-IQ1_M）が正式サポートされました。 主な改善点： ・Strix Halo：Qwen3.8のprompt処理が最大約1.8倍に高速化。DeepSeekもpromptが大幅に速くなり、F16 kernelがデフォルトに。・DGX Spark：DeepSeekのprompt処理が約2倍近く向上（短文で80-82 tok/s）。 ・Intel Arc B-series：promptが+4.6〜15.8%高速化。int8 attentionパスがデフォルトになり、2枚構成でのパイプラインも強化。 ・Linux NVIDIA：Gated DeltaNetがtensor core対応になりpromptが数%高速化。並列リクエスト時の最適化も入って、複数同時実行時のスループットが向上。 ・Linux…
Digest-only: confidence 0.55 below 0.60; incomplete methodology (context length, hardware state, measurement method).
N100 mini PC × Qwen3
Oct 31, 2026 · by @mr_mr122497
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 8192 | Q4_K_M | ollama | 12.0 tok/s | — |
| 8192 | Q4_K_M | ollama | 4.5 tok/s | — |
别急着买显卡，我把家里那台闲置的二手迷你主机改成本地推理机，跑了一个月来说说真实体感。 配置很朴素：N100 四核、DDR5 4800 16G 单条、核显 24EU、装 Debian 12。模型用 Ollama 拉 Qwen3 的 4B 和 8B，量化都选 Q4_K_M，上下文开到 8K。整机二手六百多，内存另配两百出头，待机 8W、满载 25W 上下，电费一个月不到两块钱，这账我先算清楚了才敢折腾。 速度实测：4B 大概 11 到 13 tok/s，8B 只有 4 到 5 tok/s。写代码注释、改文案、整理会议记录这些够用，但别指望它跟你流式唠嗑，吐字比你看字还慢半拍。 踩坑三个，记好了能省一晚上。第一，单通道内存是最大瓶颈，插满两条同规格实测能多出三成左右；第二，核显共享显存必须在 BIOS 里手动调大，默认值加载 8B 直接 OOM；第三，散热是原装机箱那个 4cm 小风扇，连续推理十分钟就降频，我换成 12cm…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, measurement method).
RTX 5090 × Qwen3.8-Flash-Next
Oct 31, 2026 · by @orenoa7
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | IQ3_S | basalt | — | 6950.0 tok/s |
| — | IQ3_S | basalt | 273.0 tok/s | — |
| — | IQ3_S | basalt | 283.0 tok/s | — |
| — | IQ3_S | strata | — | 5250.0 tok/s |
| — | IQ3_S | strata | 164.0 tok/s | — |
| — | IQ3_S | strata | 182.0 tok/s | — |
| — | IQ3_S | basalt | 451.0 tok/s | — |
| — | IQ3_S | basalt | 489.0 tok/s | — |
| — | IQ3_S | strata | 216.0 tok/s | — |
| — | IQ3_S | strata | 188.0 tok/s | — |
| — | IQ3_S | strata | 197.0 tok/s | — |
| — | IQ3_S | strata | 184.0 tok/s | — |
basaltとStrataをRTX5090 DDR5-3600 で速度比較してみた (QFN IQ3_S) ■ basaltって何？ StrataをベースにBlackwell GPU向けのExpertキャッシュ管理やCPU/GPU協調処理、MTPを最適化した推論エンジン basalt https://t.co/0RZEt5jbFC 実測の結果 ■ 単独 Prefill｜Prose｜Code (tok/s) ・Basalt 6950｜273｜283 ・Strata 5250｜164｜182 単独でもBasaltが3項目とも上回った prefill は+1700tok/s proseは +100tok/s codeも+100tok/s 単独に関しては想像以上にbasaltで良い結果が出た ■ 実効prefillの並列性能(2枚目)…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, context length, hardware state, measurement method).
DGX Spark x2 × DeepSeek-V4.1-Flash-1017
Oct 31, 2026 · by @NullBurn
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| 32K | — | — | — | 2082.0 tok/s |
| 32K | — | — | — | 1984.0 tok/s |
| — | — | — | 95.3 tok/s | — |
| — | — | — | 94.6 tok/s | — |
GPU clock cap alone isn't enough — cap the CPU too. On 2x DGX Spark (GB10) running DeepSeek-V4.1-Flash-1017, I capped the Cortex-X925 performance cores from 3900 to 2808 MHz and left the A725 efficiency cores alone. What changed: · 10-core CPU load: SoC peak 94°C → 74°C · prefill 32k: 2082 → 1984 tok/s (-5%) · decode x4: 95.3 → 94.6 tok/s (within noise) The GPU cap is nvidia-smi -lgc 0,2000. Why it works…
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, measurement method).
ZBook (Ryzen AI Max+ 395) × Qwen3.8-27B
Oct 31, 2026 · by @QwenAIFans · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | llama.cpp patched | 46.8 tok/s | — |
| — | — | talus | 59.9 tok/s | — |
| 32768 | — | llama.cpp patched | — | 270.0 tok/s |
| 32768 | — | talus | — | 416.0 tok/s |
💥同一台 70W 本，Qwen3.8-27B 编码约 46.8 → 59.9 tok/s 换的不是硬件。ZBook（Ryzen AI Max+ 395）上 Talus 对补丁版 llama.cpp；32K 预填充约 270 → 416。 温馨提示：作者自测，本机请复测。你会换推理引擎吗？ 来源：Hugging Face Forum👇 https://discuss.huggingface.co/t/strix-halo-tuned-rust-engine-5-models/191158 https://t.co/bOCp8HHNMB
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, hardware state, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Oct 31, 2026 · by @sgl_project
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | sglang | 943.0 tok/s | — |
No alerts, but DeepSeek V4.1 Flash with SGLang went from 654K to 1.09M tok/s per GPU in peak total throughput (cache + new input + output) on AgentX over the past four days 🤫 For ultra-low latency, we also measured 943 tok/s/user at concurrency 1 with MegaKernel. There is https://t.co/xZ5o5m2zE5
Digest-only: hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × V4.1
Oct 31, 2026 · by @teortaxesTex
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 1488.0 tok/s | — |
V4.1 at concerningly close to 1488 tokens per second, without reasoning… actually it's still on par with MiMo 2.6-Pro and Mistral 4 Large on some tasks lmao 0.7 minutes per task, 42 seconds. 1.5s end to end response (500t) https://t.co/0tCRVhcyd6
Digest-only: confidence 0.55 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
DGX Spark × qwen3.8-27b
Oct 31, 2026 · by @dhakal_ananda
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | — | 4.3 tok/s | — |
| — | — | — | 7.0 tok/s | — |
huge win for local models!! i managed to increase the token output of qwen3.8-27b by ~60% (from 4.3 tok/s to 7 tok/s) running full dense model with no tweaking is slow asf in dgx spark even if you have enough memory
Digest-only: no run with engine+version+quant+context+decode tps; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × DeepSeek V4.1 Flash
Oct 31, 2026 · by @xiangxiang103 · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | — | ollama | 202.0 tok/s | — |
都是卖 DeepSeek V4.1 Flash，这四家的路数完全不是一回事！ 数据来自官网和网友的实测 便宜的三家拼命比“谁送的额度多”。 Ollama 只拼一件事：让你少等几秒。 我的判断：真拿 AI 天天干活的人，最后都会为速度买单。 OpenCode Go 额度诚意最高，10 块给 100 块额度。 但白天高峰单价翻倍，卡顿、429 满天飞。 Command Code 也是 10 块给 100 块。 可规则改了至少 5 次，前两天还刚注销过全员 API key。 OpenDesign Go 水分最大、最不透明。 号称给 120 块，其实全天按高峰价固定扣，折算下来也就 80 块。 而且上线后为各种规则混乱道歉重置了 3 次，退款也极难。 Ollama Pro 最贵，20 块一个月，只给 60 块额度。 但测速直接飙到 202 tok/s、首字 0.3 秒，没有恶心人的 5 小时和周上限 要省钱薅羊毛选…
Digest-only: confidence 0.40 below 0.60; hardware.name missing; incomplete methodology (engine+version, quantization, context length, hardware state, measurement method).
Unspecified platform × Qwen3.8
Oct 31, 2026 · by @QwenAIFans · third-party report
| Context | Quant | Engine | Decode | Prefill |
|---|---|---|---|---|
| — | q5_1 KV cache | llama.cpp | — | 38.0 tok/s |
| — | q8_0 KV cache | llama.cpp | — | 1348.0 tok/s |
💥KV 设错，Qwen3.8 读提示慢约 35 倍 开着 Flash Attention 却用了 q5_1：不报错，注意力悄悄跑到 CPU。 换成 q8_0：预填充约 38 → 约 1348 tok/s。 温馨提示：默认只认 f16/q8_0/q4_0。你中过吗？ 来源：CheapTokens👇 https://cheaptokens.dev/en/blog/llama-cpp-slow-prompt-quantized-kv-cache/ https://t.co/h5JUCAPSt9
Digest-only: hardware.name missing; incomplete methodology (engine+version, context length, measurement method).
Also in this window
The 12 most recent posts are shown in full above; 22 posts from the same window are listed here in short form, each with its source post.
- Unspecified platform × Qwen3.8-Flash-Next — 77.2 tok/s decode (2 measurements) · @murasametech
- RTX 5090 × GLM 5.3-Flash-Maya-S — 36.0 tok/s decode (4 measurements) · @JoeGuglielmucci
- Unspecified platform × uncensored DeepSeek — 300.0 tok/s decode (1 measurement) · @silennai
- DGX Spark × Qwen3.8 Flash-Next — 83.0 tok/s decode (2 measurements) · @QwenAIFans
- DGX Spark × Qwen3.8 Flash-Next — 83.0 tok/s decode (2 measurements) · @filicroval
- Unspecified platform — 60.0 tok/s decode (1 measurement) · @yuhasbeentaken
- AMD Gorgon x2 × DeepSeek V4.1 Flash 1017 — 72.0 tok/s decode (1 measurement) · @Italianclownz
- Mac Studio M5 Ultra 512GB × GLM-5.3 — 75.0 tok/s decode (4 measurements) · @tomgreenwald
- M5 Max × Qwen3.6 35B — 150.0 tok/s decode (3 measurements) · @rapidmlx
- RTX 3090 x2 × Qwen3.8-27B INT4 — 123.0 tok/s decode (2 measurements) · @QwenAIFans
- RTX 3090 x4 × DeepSeek V4.1 Flash 1017 — 39.7 tok/s decode (1 measurement) · @superalesha
- DGX Spark × DeepSeek V4.1 Flash — 122.1 tok/s decode (3 measurements) · @tenhkspark
- DGX Spark × DeepSeek V4.1 Flash 1017 SAGE-EXL3 — 35.5 tok/s decode (1 measurement) · @ViC305
- DDR3 Xeon x2 + modded mining cards × Qwen3.8 Flash Next 180B — 75.0 tok/s decode (1 measurement) · @QwenAIFans
- RTX 2080 × Qwen Next 125B — 845.0 tok/s prefill (1 measurement) · @riccardogiorato
- RTX PRO 4500 Blackwell × Qwen3.8-27B — 153.0 tok/s decode (1 measurement) · @lasse_ai
- NVIDIA GPU × Ingot — 250.0 tok/s decode (1 measurement) · @ingot_wthrboss
- DGX Spark x4 × DeepSeek-V4.1-Flash-1017 — 7601.0 tok/s prefill (6 measurements) · @tenhkspark
- RTX 5090 × DeepSeek V4.1 Flash — 40.0 tok/s decode (1 measurement) · @pratikg
- CMP 50HX 20GB x2 × Qwen 3.8 Flash Next — 75.1 tok/s decode (2 measurements) · @VRAMCalculator
- Apple M5 Max 128GB × DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw — 161.0 tok/s decode (10 measurements) · @Beamsters1
- RTX 3060 × Qwen3.8 Flash Next — 68.0 tok/s decode (1 measurement) · @QwenAIFans
How to read this page
This digest is an automated transcription of public posts. Runs with a stated engine build, quantization, context length, hardware state and measurement method are published straight to the benchmark board with a link to their source; the rest stay here. These are author-reported numbers, not an independent DiggTop re-measurement: every row carries its provenance so you can check the original, and a number that turns out to be misreported is corrected or removed. Our own testing follows the methodology. Spotted an error? Reply on the source post or submit a corrected run at /submit.
Sources
- @yuracafe_ — Oct 31, 2026
- @jun1228909 — Oct 31, 2026
- @murasametech — Oct 31, 2026
- @mr_mr122497 — Oct 31, 2026
- @orenoa7 — Oct 31, 2026
- @NullBurn — Oct 31, 2026
- @QwenAIFans — Oct 31, 2026
- @sgl_project — Oct 31, 2026
- @teortaxesTex — Oct 31, 2026
- @dhakal_ananda — Oct 31, 2026
- @xiangxiang103 — Oct 31, 2026
- @QwenAIFans — Oct 31, 2026
- @murasametech — Oct 31, 2026
- @JoeGuglielmucci — Oct 30, 2026
- @silennai — Oct 30, 2026
- @QwenAIFans — Oct 30, 2026
- @filicroval — Oct 30, 2026
- @yuhasbeentaken — Oct 30, 2026
- @Italianclownz — Oct 30, 2026
- @tomgreenwald — Oct 30, 2026
- @rapidmlx — Oct 30, 2026
- @QwenAIFans — Oct 30, 2026
- @superalesha — Oct 30, 2026
- @tenhkspark — Oct 30, 2026
- @ViC305 — Oct 30, 2026
- @QwenAIFans — Oct 30, 2026
- @riccardogiorato — Oct 30, 2026
- @lasse_ai — Oct 30, 2026
- @ingot_wthrboss — Oct 30, 2026
- @tenhkspark — Oct 30, 2026
- @pratikg — Oct 30, 2026
- @VRAMCalculator — Oct 30, 2026
- @Beamsters1 — Oct 30, 2026
- @QwenAIFans — Oct 30, 2026
FAQ
Are the numbers in this digest sourced?
They are transcribed from public X posts and the repositories those posts link to, with attribution and a source link on every entry. Rows that state the full methodology are published to the benchmark board automatically, so treat them as author-reported until a dispute is raised: the source link on every row is the check.
What does it take for a row to reach the sourced tier?
Engine and exact version, quantization, context length, hardware state and a described measurement method. Posts that state all five are published to the benchmark board as soon as the collector reads them, each row carrying its source link — nothing waits in a human review queue. Anything missing a condition stays in this digest.
Where does the data come from?
Public X posts about local-model inference measured on owned hardware, discovered by keyword search every two hours and linked back to the original thread. The digest is regenerated daily.